From: Bijan Tabatabai <btabatabai@wisc.edu>

The man page for pkey_alloc(2) specifies that it should return -1 with
the errno set to ENOSPC when pkeys are not supported [1]. However, on
x86 pkey_alloc() sets errno to EINVAL when called for the first time
on a CPU that does not support pkeys.

The root cause of this is the x86 implementation of mm_pkey_alloc() not
directly checking if pkeys are supported. It only checks if all the
pkeys have been allocated by comparing the allocation map against
all_pkeys_mask. When OSPKE is not enabled, init_new_context() skips the
initialization of the allocation map, leaving it as 0, while
all_pkeys_mask is 1. mm_pkey_alloc() interprets this as there being a
pkey available and it returns pkey 0. Then, pkey_alloc() fails with
-EINVAL from arch_set_user_pkey_access() instead of returning -ENOSPC.
Subsequent calls to pkey_alloc() do return -ENOSPC because pkey 0 is
left marked as allocated.

Change mm_pkey_alloc() to directly check if OSPKE is enabled, and
return -1 if it is not, which causes pkey_alloc() to return -ENOSPC. The
arm64 and powerpc implementations of mm_pkey_alloc() already do this
check.

[1] https://man7.org/linux/man-pages/man2/pkey_alloc.2.html

Fixes: e8c24d3a23a4 ("x86/pkeys: Allocation/free syscalls")
Signed-off-by: Bijan Tabatabai <btabatabai@wisc.edu>
---
 arch/x86/include/asm/pkeys.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/arch/x86/include/asm/pkeys.h b/arch/x86/include/asm/pkeys.h
index 06ed2cd2592e..fbb15f6666f1 100644
--- a/arch/x86/include/asm/pkeys.h
+++ b/arch/x86/include/asm/pkeys.h
@@ -88,6 +88,9 @@ int mm_pkey_alloc(struct mm_struct *mm)
 	u16 all_pkeys_mask = ((1U << arch_max_pkey()) - 1);
 	int ret;
 
+	if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
+		return -1;
+
 	/*
 	 * Are we out of pkeys?  We must handle this specially
 	 * because ffz() behavior is undefined if there are no
-- 
2.34.1

On 7/16/26 15:06, Bijan Tabatabai wrote:
> Change mm_pkey_alloc() to directly check if OSPKE is enabled, and
> return -1 if it is not, which causes pkey_alloc() to return -ENOSPC. The
> arm64 and powerpc implementations of mm_pkey_alloc() already do this
> check.

Thanks for the patch! I made one tweak which was to use
arch_pkeys_enabled() instead of the cpu_feature_enabled() check. That
makes the code darn close to a perfect copy of the ARM code at least.

It would be really nice if someone could go consolidate all the copying
and pasting between the architectures there some day.